cop

• Version: 0.1-0
• URL: https://github.com/RBigData/cop
• License:BSD 2-Clause
• Author: Drew Schmidt

Some custom reduction operations for data distributed over MPI.

Installation

The development version is maintained on GitHub. Because the package uses git submodules, you can not use any of the install_github() functions. Instead you can do this:

source("https://gist.githubusercontent.com/wrathematics/ccf6bf366279e099563e69e56b4fde59/raw/6daf038ca56f4f531f631f80a54d8d9aa70b4bbb/ighwsm.r")
ighwsm("RBigData/cop")

API

Dense, numeric matrix reductions:

Method	Explanation
cop_allreduce()	Reduction that accepts a custom operation (R function). The function has some strong caveats; see ?cop::cop_allreduce for details.
cop_reduce()
qr_allreduce()	Reduction where each process owns the R matrix of a QR decomposition, and the reduction is (conceptually) op = function(a, b) qr.R(qr(a, b)).
qr_reduce()

Sparse matrix reductions:

Method	Explanation
spadd_allreduce()	Reduction where each process owns a sparse matrix (dgCMatrix from the Matrix package), and the reduction sums all of the matrices.
spadd_reduce()

The difference between the _reduce() and the _allreduce() variants is that with the former, only the process specified by the root argument receives the return.

The package also has some helper utilities:

Method	Explanation
mpi_cat()	Helper cat() function.
mpi_print()	Helper print() function.

These are similar to pbdMPI::comm.cat() and pbdMPI::comm.print(), although here printing is guaranteed to occur in rank order.

Examples

For the sake of example, we can build a matrix addition custom reducer:

suppressMessages(library(cop))
x=matrix(1:4, 2) +10*comm.rank()
add=function(a, b) a+bout= cop_allreduce(x, op=add, commutative=TRUE)
mpi_cat("\nfunction:\n", quiet=TRUE)
mpi_print(add)
mpi_cat("\noutput:\n", quiet=TRUE)
mpi_print(out)
finalize()

Most of this is just printing and boilerplate, but even so it's still pretty simple. If we save this as add.r and run it via mpirun -np 3 Rscript add.r, we see:

function:
(rank=0)
function (a, b) a + b
output:
(rank=0)
[,1] [,2]
[1,] 33 39
[2,] 36 42

With three ranks, conceptually this is the same as calculating:

add=function(a, b) a+b
add(x, add(x+10, x+20))
## [,1] [,2]## [1,] 33 39## [2,] 36 42

As a more substantive example, we can build a matrix product reducer.

suppressMessages(library(cop))
x=matrix(1:4, 2) +10*comm.rank()
mult=function(a, b) a%*%bout= cop_allreduce(x, op=mult, commutative=FALSE)
mpi_cat("\nfunction:\n", quiet=TRUE)
mpi_print(mult)
mpi_cat("\noutput:\n", quiet=TRUE)
mpi_print(out)
finalize()

As before, most of the code is just boilerplate/printing. Really we only change the reducing function from the above. Note that we have to specify that the operation is not commutative.

If we save this as mult.r and run it via mpirun -np 3 Rscript mult.r, we see:

function:
(rank=0)
function (a, b) a %*% b
output:
(rank=0)
[,1] [,2]
[1,] 2197 2401
[2,] 3274 3578

With three ranks, conceptually this is the same as calculating:

x=matrix(1:4, 2)
mult=function(a, b) a%*%b
mult(x, mult(x+10, x+20))
## [,1] [,2]## [1,] 2197 2401## [2,] 3274 3578

You can find more examples in the inst/examples/ directory of the package source, or in examples/ of the installed package, located at system.file("examples", package="cop").